YOLO9000:

Better, Faster, Stronger

Joseph Redmon∗×, Ali Farhadi∗†×

University of Washington∗, Allen Institute for AI†, XNOR.ai×

http://pjreddie.com/yolo9000/

Abstract

We introduce YOLO9000, a state-of-the-art, real-time

object detection system that can detect over 9000 object

categories. First we propose various improvements to the

YOLO detection method, both novel and drawn from prior

work. The improved model, YOLOv2, is state-of-the-art on

standard detection tasks like PASCAL VOC and COCO. Us-

ing a novel, multi-scale training method the same YOLOv2

model can run at varying sizes, offering an easy tradeoff

between speed and accuracy. At 67 FPS, YOLOv2 gets

76.8 mAP on VOC 2007. At 40 FPS, YOLOv2 gets 78.6

mAP, outperforming state-of-the-art methods like Faster R-

CNN with ResNet and SSD while still running significantly

faster. Finally we propose a method to jointly train on ob-

ject detection and classification. Using this method we train

YOLO9000 simultaneously on the COCO detection dataset

and the ImageNet classification dataset. Our joint training

allows YOLO9000 to predict detections for object classes

that don’t have labelled detection data. We validate our

approach on the ImageNet detection task. YOLO9000 gets

19.7 mAP on the ImageNet detection validation set despite

only having detection data for 44 of the 200 classes. On

the 156 classes not in COCO, YOLO9000 gets 16.0 mAP.

YOLO9000 predicts detections for more than 9000 different

object categories, all in real-time.

1. Introduction

General purpose object detection should be fast, accu-

rate, and able to recognize a wide variety of objects. Since

the introduction of neural networks, detection frameworks

have become increasingly fast and accurate. However, most

detection methods are still constrained to a small set of ob-

jects.

Current object detection datasets are limited compared

to datasets for other tasks like classification and tagging.

The most common detection datasets contain thousands to

hundreds of thousands of images with dozens to hundreds

of tags [3] [10] [2]. Classification datasets have millions

of images with tens or hundreds of thousands of categories

[20] [2].

We would like detection to scale to level of object clas-

sification. However, labelling images for detection is far

more expensive than labelling for classification or tagging

(tags are often user-supplied for free). Thus we are unlikely

to see detection datasets on the same scale as classification

Figure 1: YOLO9000. YOLO9000 can detect a wide variety ofobject classes in real-time.

7263

http://pjreddie.com/yolo9000/

datasets in the near future.

We propose a new method to harness the large amount

of classification data we already have and use it to expand

the scope of current detection systems. Our method uses a

hierarchical view of object classification that allows us to

combine distinct datasets together.

We also propose a joint training algorithm that allows

us to train object detectors on both detection and classifica-

tion data. Our method leverages labeled detection images to

learn to precisely localize objects while it uses classification

images to increase its vocabulary and robustness.

Using this method we train YOLO9000, a real-time ob-

ject detector that can detect over 9000 different object cat-

egories. First we improve upon the base YOLO detection

system to produce YOLOv2, a state-of-the-art, real-time

detector. Then we use our dataset combination method

and joint training algorithm to train a model on more than

9000 classes from ImageNet as well as detection data from

COCO.

All of our code and pre-trained models are available on-

line at http://pjreddie.com/yolo9000/.

2. Better

YOLO suffers from a variety of shortcomings relative to

state-of-the-art detection systems. Error analysis of YOLO

compared to Fast R-CNN shows that YOLO makes a sig-

nificant number of localization errors. Furthermore, YOLO

has relatively low recall compared to region proposal-based

methods. Thus we focus mainly on improving recall and

localization while maintaining classification accuracy.

Computer vision generally trends towards larger, deeper

networks [6] [18] [17]. Better performance often hinges on

training larger networks or ensembling multiple models to-

gether. However, with YOLOv2 we want a more accurate

detector that is still fast. Instead of scaling up our network,

we simplify the network and then make the representation

easier to learn. We pool a variety of ideas from past work

with our own novel concepts to improve YOLO’s perfor-

mance. A summary of results can be found in Table 2.

Batch Normalization. Batch normalization leads to sig-

nificant improvements in convergence while eliminating the

need for other forms of regularization [7]. By adding batch

normalization on all of the convolutional layers in YOLO

we get more than 2% improvement in mAP. Batch normal-

ization also helps regularize the model. With batch nor-

malization we can remove dropout from the model without

overfitting.

High Resolution Classifier. All state-of-the-art detec-

tion methods use classifier pre-trained on ImageNet [16].

Starting with AlexNet most classifiers operate on input im-

ages smaller than 256× 256 [8]. The original YOLO trainsthe classifier network at 224 × 224 and increases the reso-lution to 448 for detection. This means the network has tosimultaneously switch to learning object detection and ad-

just to the new input resolution.

For YOLOv2 we first fine tune the classification network

at the full 448× 448 resolution for 10 epochs on ImageNet.This gives the network time to adjust its filters to work better

on higher resolution input. We then fine tune the resulting

network on detection. This high resolution classification

network gives us an increase of almost 4% mAP.

Convolutional With Anchor Boxes. YOLO predicts

the coordinates of bounding boxes directly using fully con-

nected layers on top of the convolutional feature extractor.

Instead of predicting coordinates directly Faster R-CNN

predicts bounding boxes using hand-picked priors [15]. Us-

ing only convolutional layers the region proposal network

(RPN) in Faster R-CNN predicts offsets and confidences for

anchor boxes. Since the prediction layer is convolutional,

the RPN predicts these offsets at every location in a feature

map. Predicting offsets instead of coordinates simplifies the

problem and makes it easier for the network to learn.

We remove the fully connected layers from YOLO and

use anchor boxes to predict bounding boxes. First we

eliminate one pooling layer to make the output of the net-

work’s convolutional layers higher resolution. We also

shrink the network to operate on 416 input images insteadof 448×448. We do this because we want an odd number oflocations in our feature map so there is a single center cell.

Objects, especially large objects, tend to occupy the center

of the image so it’s good to have a single location right at

the center to predict these objects instead of four locations

that are all nearby. YOLO’s convolutional layers downsam-

ple the image by a factor of 32 so by using an input image

of 416 we get an output feature map of 13× 13.When we move to anchor boxes we also decouple the

class prediction mechanism from the spatial location and

instead predict class and objectness for every anchor box.

Following YOLO, the objectness prediction still predicts

the IOU of the ground truth and the proposed box and the

class predictions predict the conditional probability of that

class given that there is an object.

Using anchor boxes we get a small decrease in accuracy.

YOLO only predicts 98 boxes per image but with anchor

boxes our model predicts more than a thousand. Without

anchor boxes our intermediate model gets 69.5 mAP with arecall of 81%. With anchor boxes our model gets 69.2 mAPwith a recall of 88%. Even though the mAP decreases, theincrease in recall means that our model has more room to

improve.

Dimension Clusters. We encounter two issues with an-

chor boxes when using them with YOLO. The first is that

the box dimensions are hand picked. The network can learn

to adjust the boxes appropriately but if we pick better priors

for the network to start with we can make it easier for the

network to learn to predict good detections.

Instead of choosing priors by hand, we run k-means

clustering on the training set bounding boxes to automat-

ically find good priors. If we use standard k-means with

7264

http://pjreddie.com/yolo9000/

0

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

COCO

# Clusters

Avg IOU

0.75

VOC 2007

Figure 2: Clustering box dimensions on VOC and COCO. Werun k-means clustering on the dimensions of bounding boxes to get

good priors for our model. The left image shows the average IOU

we get with various choices for k. k = 5 gives a good tradeoff for

recall vs. complexity of the model. The right image shows the rel-

ative centroids for VOC and COCO. COCO has greater variation

in size than VOC.

Euclidean distance larger boxes generate more error than

smaller boxes. However, what we really want are priors

that lead to good IOU scores, which is independent of the

size of the box. Thus for our distance metric we use:

d(box, centroid) = 1− IOU(box, centroid)

We run k-means for various values of k and plot the av-

erage IOU with closest centroid, see Figure 2. We choose

k = 5 as a good tradeoff between model complexity andhigh recall. The cluster centroids are significantly different

than hand-picked anchor boxes. There are fewer short, wide

boxes and more tall, thin boxes.

We compare the average IOU to closest prior of our clus-

tering strategy and the hand-picked anchor boxes in Table 1.

At only 5 priors the centroids perform similarly to 9 anchor

boxes with an average IOU of 61.0 compared to 60.9. If

we use 9 centroids we see a much higher average IOU. This

indicates that using k-means to generate our bounding box

starts the model off with a better representation and makes

the task easier to learn.

Box Generation # Avg IOUCluster SSE 5 58.7

Cluster IOU 5 61.0

Anchor Boxes [15] 9 60.9

Cluster IOU 9 67.2

Table 1: Average IOU of boxes to closest priors on VOC 2007.The average IOU of objects on VOC 2007 to their closest, unmod-

ified prior using different generation methods. Clustering gives

much better results than using hand-picked priors.

Direct location prediction. When using anchor boxes

with YOLO we encounter a second issue: model instability,

especially during early iterations. Most of the instability

comes from predicting the (x, y) locations for the box. Inregion proposal networks the network predicts values tx and

ty and the (x, y) center coordinates are calculated as:

x = (tx ∗ wa)− xa

y = (ty ∗ ha)− ya

For example, a prediction of tx = 1 would shift the boxto the right by the width of the anchor box, a prediction of

tx = −1 would shift it to the left by the same amount.This formulation is unconstrained so any anchor box can

end up at any point in the image, regardless of what loca-

tion predicted the box. With random initialization the model

takes a long time to stabilize to predicting sensible offsets.

Instead of predicting offsets we follow the approach of

YOLO and predict location coordinates relative to the loca-

tion of the grid cell. This bounds the ground truth to fall

between 0 and 1. We use a logistic activation to constrainthe network’s predictions to fall in this range.

The network predicts 5 bounding boxes at each cell in

the output feature map. The network predicts 5 coordinates

for each bounding box, tx, ty , tw, th, and to. If the cell is

offset from the top left corner of the image by (cx, cy) andthe bounding box prior has width and height pw, ph, then

the predictions correspond to:

bx = σ(tx) + cx

by = σ(ty) + cy

bw = pwetw

bh = pheth

Pr(object) ∗ IOU(b, object) = σ(to)

Since we constrain the location prediction the

parametrization is easier to learn, making the network

more stable. Using dimension clusters along with directly

predicting the bounding box center location improves

YOLO by almost 5% over the version with anchor boxes.

Fine-Grained Features.This modified YOLO predicts

detections on a 13 × 13 feature map. While this is suffi-cient for large objects, it may benefit from finer grained fea-

tures for localizing smaller objects. Faster R-CNN and SSD

both run their proposal networks at various feature maps in

the network to get a range of resolutions. We take a differ-

ent approach, simply adding a passthrough layer that brings

features from an earlier layer at 26× 26 resolution.The passthrough layer concatenates the higher resolution

features with the low resolution features by stacking adja-

cent features into different channels instead of spatial lo-

cations, similar to the identity mappings in ResNet. This

turns the 26× 26× 512 feature map into a 13× 13× 2048

7265

σ(tx)

σ(ty)

pw

ph bh

bw

bw=p

we

bh=p

he

cx

cy

bx=σ(t

x)+c

x

by=σ(t

y)+c

ytw

th

Figure 3: Bounding boxes with dimension priors and locationprediction. We predict the width and height of the box as offsets

from cluster centroids. We predict the center coordinates of the

box relative to the location of filter application using a sigmoid

function.

feature map, which can be concatenated with the original

features. Our detector runs on top of this expanded feature

map so that it has access to fine grained features. This gives

a modest 1% performance increase.

Multi-Scale Training. The original YOLO uses an input

resolution of 448× 448. With the addition of anchor boxeswe changed the resolution to 416×416. However, since ourmodel only uses convolutional and pooling layers it can be

resized on the fly. We want YOLOv2 to be robust to running

on images of different sizes so we train this into the model.

Instead of fixing the input image size we change the net-

work every few iterations. Every 10 batches our network

randomly chooses new image dimensions. Since our model

downsamples by a factor of 32, we pull from the following

multiples of 32: {320, 352, ..., 608}. Thus the smallest op-tion is 320×320 and the largest is 608×608. We resize thenetwork to that dimension and continue training.

This regime forces the network to learn to predict well

across a variety of input dimensions. This means the same

network can predict detections at different resolutions. The

network runs faster at smaller sizes so YOLOv2 offers an

easy tradeoff between speed and accuracy.

At low resolutions YOLOv2 operates as a cheap, fairly

accurate detector. At 288× 288 it runs at more than 90 FPSwith mAP almost as good as Fast R-CNN. This makes it

ideal for smaller GPUs, high framerate video, or multiple

video streams.

At high resolution YOLOv2 is a state-of-the-art detector

with 78.6 mAP on VOC 2007 while still operating above

real-time speeds. See Table 3 for a comparison of YOLOv2

with other frameworks on VOC 2007. Figure 4

Further Experiments. We train YOLOv2 for detection

Mean Average Precision

Frames Per Second

R-CNN

YOLO

Fast R-CNN

Faster R-CNN

Faster R-CNNResnet

SSD512

SSD300

YOLOv2

80

70

60

0 50 10030

Figure 4: Accuracy and speed on VOC 2007.

on VOC 2012. Table 4 shows the comparative performance

of YOLOv2 versus other state-of-the-art detection systems.

YOLOv2 achieves 73.4 mAP while running far faster than

other methods. We also train on COCO, see Table 5. On the

VOC metric (IOU = .5) YOLOv2 gets 44.0 mAP, compara-

ble to SSD and Faster R-CNN.

3. Faster

We want detection to be accurate but we also want it to be

fast. Most applications for detection, like robotics or self-

driving cars, rely on low latency predictions. In order to

maximize performance we design YOLOv2 to be fast from

the ground up.

Most detection frameworks rely on VGG-16 as the base

feature extractor [17]. VGG-16 is a powerful, accurate clas-

sification network but it is needlessly complex. The con-

volutional layers of VGG-16 require 30.69 billion floating

point operations for a single pass over a single image at

224× 224 resolution.The YOLO framework uses a custom network based on

the Googlenet architecture [19]. This network is faster than

VGG-16, only using 8.52 billion operations for a forward

pass. However, it’s accuracy is slightly worse than VGG-

16. For single-crop, top-5 accuracy at 224 × 224, YOLO’scustom model gets 88.0% ImageNet compared to 90.0% for

VGG-16.

Darknet-19. We propose a new classification model to

be used as the base of YOLOv2. Our model builds off of

prior work on network design as well as common knowl-

edge in the field. Similar to the VGG models we use mostly

3 × 3 filters and double the number of channels after ev-ery pooling step [17]. Following the work on Network in

Network (NIN) we use global average pooling to make pre-

dictions as well as 1× 1 filters to compress the feature rep-resentation between 3 × 3 convolutions [9]. We use batchnormalization to stabilize training, speed up convergence,

7266

YOLO YOLOv2

batch norm? X X X X X X X X

hi-res classifier? X X X X X X X

convolutional? X X X X X X

anchor boxes? X X

new network? X X X X X

dimension priors? X X X X

location prediction? X X X X

passthrough? X X X

multi-scale? X X

hi-res detector? X

VOC2007 mAP 63.4 65.8 69.5 69.2 69.6 74.4 75.4 76.8 78.6

Table 2: The path from YOLO to YOLOv2. Most of the listed design decisions lead to significant increases in mAP. Two

exceptions are switching to a fully convolutional network with anchor boxes and using the new network. Switching to the

anchor box style approach increased recall without changing mAP while using the new network cut computation by 33%.

Detection Frameworks Train mAP FPS

Fast R-CNN [5] 2007+2012 70.0 0.5

Faster R-CNN VGG-16[15] 2007+2012 73.2 7

Faster R-CNN ResNet[6] 2007+2012 76.4 5

YOLO [14] 2007+2012 63.4 45

SSD300 [11] 2007+2012 74.3 46

SSD500 [11] 2007+2012 76.8 19

YOLOv2 288× 288 2007+2012 69.0 91YOLOv2 352× 352 2007+2012 73.7 81YOLOv2 416× 416 2007+2012 76.8 67YOLOv2 480× 480 2007+2012 77.8 59YOLOv2 544× 544 2007+2012 78.6 40

Table 3: Detection frameworks on PASCAL VOC 2007.YOLOv2 is faster and more accurate than prior detection meth-

ods. It can also run at different resolutions for an easy tradeoff

between speed and accuracy. Each YOLOv2 entry is actually the

same trained model with the same weights, just evaluated at a dif-

ferent size. All timing information is on a Geforce GTX Titan X

(original, not Pascal model).

and regularize the model [7].

Our final model, called Darknet-19, has 19 convolutional

layers and 5 maxpooling layers. For a full description see

Table 6. Darknet-19 only requires 5.58 billion operations

to process an image yet achieves 72.9% top-1 accuracy and91.2% top-5 accuracy on ImageNet.

Training for classification. We train the network on

the standard ImageNet 1000 class classification dataset for

160 epochs using stochastic gradient descent with a starting

learning rate of 0.1, polynomial rate decay with a power of4, weight decay of 0.0005 and momentum of 0.9 using theDarknet neural network framework [13]. During training

we use standard data augmentation tricks including random

crops, rotations, and hue, saturation, and exposure shifts.

As discussed above, after our initial training on images

at 224× 224 we fine tune our network at a larger size, 448.For this fine tuning we train with the above parameters but

for only 10 epochs and starting at a learning rate of 10−3. Atthis higher resolution our network achieves a top-1 accuracy

of 76.5% and a top-5 accuracy of 93.3%.Training for detection. We modify this network for de-

tection by removing the last convolutional layer and instead

adding on three 3 × 3 convolutional layers with 1024 fil-ters each followed by a final 1× 1 convolutional layer withthe number of outputs we need for detection. For VOC we

predict 5 boxes with 5 coordinates each and 20 classes per

box so 125 filters. We also add a passthrough layer from the

final 3 × 3 × 512 layer to the second to last convolutionallayer so that our model can use fine grain features.

We train the network for 160 epochs with a starting

learning rate of 10−3, dividing it by 10 at 60 and 90 epochs.We use a weight decay of 0.0005 and momentum of 0.9.We use a similar data augmentation to YOLO and SSD with

random crops, color shifting, etc. We use the same training

strategy on COCO and VOC.

4. Stronger

We propose a mechanism for jointly training on classi-

fication and detection data. Our method uses images la-

belled for detection to learn detection-specific information

like bounding box coordinate prediction and objectness as

well as how to classify common objects. It uses images with

only class labels to expand the number of categories it can

detect.

During training we mix images from both detection and

classification datasets. When our network sees an image

labelled for detection we can backpropagate based on the

full YOLOv2 loss function. When it sees a classification

image we only backpropagate loss from the classification-

specific parts of the architecture.

This approach presents a few challenges. Detection

datasets have only common objects and general labels, like

7267

Method data mAP aero bike bird boat bottle bus car cat chair cow table dog horse mbike person plant sheep sofa train tv

Fast R-CNN [5] 07++12 68.4 82.3 78.4 70.8 52.3 38.7 77.8 71.6 89.3 44.2 73.0 55.0 87.5 80.5 80.8 72.0 35.1 68.3 65.7 80.4 64.2Faster R-CNN [15] 07++12 70.4 84.9 79.8 74.3 53.9 49.8 77.5 75.9 88.5 45.6 77.1 55.3 86.9 81.7 80.9 79.6 40.1 72.6 60.9 81.2 61.5YOLO [14] 07++12 57.9 77.0 67.2 57.7 38.3 22.7 68.3 55.9 81.4 36.2 60.8 48.5 77.2 72.3 71.3 63.5 28.9 52.2 54.8 73.9 50.8SSD300 [11] 07++12 72.4 85.6 80.1 70.5 57.6 46.2 79.4 76.1 89.2 53.0 77.0 60.8 87.0 83.1 82.3 79.4 45.9 75.9 69.5 81.9 67.5SSD512 [11] 07++12 74.9 87.4 82.3 75.8 59.0 52.6 81.7 81.5 90.0 55.4 79.0 59.8 88.4 84.3 84.7 83.3 50.2 78.0 66.3 86.3 72.0ResNet [6] 07++12 73.8 86.5 81.6 77.2 58.0 51.0 78.6 76.6 93.2 48.6 80.4 59.0 92.1 85.3 84.8 80.7 48.1 77.3 66.5 84.7 65.6

YOLOv2 544 07++12 73.4 86.3 82.0 74.8 59.2 51.8 79.8 76.5 90.6 52.1 78.2 58.5 89.3 82.5 83.4 81.3 49.1 77.2 62.4 83.8 68.7

Table 4: PASCAL VOC2012 test detection results. YOLOv2 performs on par with state-of-the-art detectors like Faster

R-CNN with ResNet and SSD512 and is 2− 10× faster.

0.5:0.95 0.5 0.75 S M L 1 10 100 S M L

Fast R-CNN [5] train 19.7 35.9 - - - - - - - - - -

Fast R-CNN[1] train 20.5 39.9 19.4 4.1 20.0 35.8 21.3 29.5 30.1 7.3 32.1 52.0

Faster R-CNN[15] trainval 21.9 42.7 - - - - - - - - - -

ION [1] train 23.6 43.2 23.6 6.4 24.1 38.3 23.2 32.7 33.5 10.1 37.7 53.6

Faster R-CNN[10] trainval 24.2 45.3 23.5 7.7 26.4 37.1 23.8 34.0 34.6 12.0 38.5 54.4

SSD300 [11] trainval35k 23.2 41.2 23.4 5.3 23.2 39.6 22.5 33.2 35.3 9.6 37.6 56.5

SSD512 [11] trainval35k 26.8 46.5 27.8 9.0 28.9 41.9 24.8 37.5 39.8 14.0 43.5 59.0

YOLOv2 [11] trainval35k 21.6 44.0 19.2 5.0 22.4 35.5 20.7 31.6 33.3 9.8 36.5 54.4

Table 5: Results on COCO test-dev2015. Table adapted from [11]

Type Filters Size/Stride Output

Convolutional 32 3 × 3 224 × 224Maxpool 2 × 2/2 112 × 112

Convolutional 64 3 × 3 112 × 112Maxpool 2 × 2/2 56 × 56

Convolutional 128 3 × 3 56 × 56Convolutional 64 1 × 1 56 × 56Convolutional 128 3 × 3 56 × 56

Maxpool 2 × 2/2 28 × 28Convolutional 256 3 × 3 28 × 28Convolutional 128 1 × 1 28 × 28Convolutional 256 3 × 3 28 × 28

Maxpool 2 × 2/2 14 × 14Convolutional 512 3 × 3 14 × 14Convolutional 256 1 × 1 14 × 14Convolutional 512 3 × 3 14 × 14Convolutional 256 1 × 1 14 × 14Convolutional 512 3 × 3 14 × 14

Maxpool 2 × 2/2 7 × 7Convolutional 1024 3 × 3 7 × 7Convolutional 512 1 × 1 7 × 7Convolutional 1024 3 × 3 7 × 7Convolutional 512 1 × 1 7 × 7Convolutional 1024 3 × 3 7 × 7

Convolutional 1000 1 × 1 7 × 7Avgpool Global 1000Softmax

Table 6: Darknet-19.

“dog” or “boat”. Classification datasets have a much wider

and deeper range of labels. ImageNet has more than a hun-

dred breeds of dog, including “Norfolk terrier”, “Yorkshire

terrier”, and “Bedlington terrier”. If we want to train on

both datasets we need a coherent way to merge these labels.

Most approaches to classification use a softmax layer

across all the possible categories to compute the final prob-

ability distribution. Using a softmax assumes the classes

are mutually exclusive. This presents problems for combin-

ing datasets, for example you would not want to combine

ImageNet and COCO using this model because the classes

“Norfolk terrier” and “dog” are not mutually exclusive.

We could instead use a multi-label model to combine the

datasets which does not assume mutual exclusion. This ap-

proach ignores all the structure we do know about the data,

for example that all of the COCO classes are mutually ex-

clusive.

Hierarchical classification. ImageNet labels are pulled

from WordNet, a language database that structures concepts

and how they relate [12]. In WordNet, “Norfolk terrier” and

“Yorkshire terrier” are both hyponyms of “terrier” which is

a type of “hunting dog”, which is a type of “dog”, which is

a “canine”, etc. Most approaches to classification assume a

flat structure to the labels however for combining datasets,

structure is exactly what we need.

WordNet is structured as a directed graph, not a tree, be-

cause language is complex. For example a “dog” is both

a type of “canine” and a type of “domestic animal” which

are both synsets in WordNet. Instead of using the full graph

structure, we simplify the problem by building a hierarchi-

cal tree from the concepts in ImageNet.

To build this tree we examine the visual nouns in Ima-

geNet and look at their paths through the WordNet graph to

the root node, in this case “physical object”. Many synsets

only have one path through the graph so first we add all of

those paths to our tree. Then we iteratively examine the

concepts we have left and add the paths that grow the tree

by as little as possible. So if a concept has two paths to the

root and one path would add three edges to our tree and the

other would only add one edge, we choose the shorter path.

The final result is WordTree, a hierarchical model of vi-

sual concepts. To perform classification with WordTree we

predict conditional probabilities at every node for the prob-

7268

ability of each hyponym of that synset given that synset. For

example, at the “terrier” node we predict:

Pr(Norfolk terrier|terrier)

Pr(Yorkshire terrier|terrier)

Pr(Bedlington terrier|terrier)

...

If we want to compute the absolute probability for a par-

ticular node we simply follow the path through the tree to

the root node and multiply to conditional probabilities. So

if we want to know if a picture is of a Norfolk terrier we

compute:

Pr(Norfolk terrier) = Pr(Norfolk terrier|terrier)

∗Pr(terrier|hunting dog)

∗ . . .∗

∗Pr(mammal|Pr(animal)

∗Pr(animal|physical object)

For classification purposes we assume that the the image

contains an object: Pr(physical object) = 1.To validate this approach we train the Darknet-19 model

on WordTree built using the 1000 class ImageNet. To build

WordTree1k we add in all of the intermediate nodes which

expands the label space from 1000 to 1369. During training

we propagate ground truth labels up the tree so that if an im-

age is labelled as a “Norfolk terrier” it also gets labelled as

a “dog” and a “mammal”, etc. To compute the conditional

probabilities our model predicts a vector of 1369 values and

we compute the softmax over all sysnsets that are hyponyms

of the same concept, see Figure 5.

Using the same training parameters as before, our hi-

erarchical Darknet-19 achieves 71.9% top-1 accuracy and90.4% top-5 accuracy. Despite adding 369 additional con-cepts and having our network predict a tree structure our ac-

curacy only drops marginally. Performing classification in

this manner also has some benefits. Performance degrades

gracefully on new or unknown object categories. For exam-

ple, if the network sees a picture of a dog but is uncertain

what type of dog it is, it will still predict “dog” with high

confidence but have lower confidences spread out among

the hyponyms.

This formulation also works for detection. Now, in-

stead of assuming every image has an object, we use

YOLOv2’s objectness predictor to give us the value of

Pr(physical object). The detector predicts a bounding boxand the tree of probabilities. We traverse the tree down, tak-

ing the highest confidence path at every split until we reach

some threshold and we predict that object class.

...

kit fox

English setter

Siberian husky

Australian terrier

English springer

grey whale

lesser panda

Egyptian cat

ibexPersian cat

cougar

rubber eraser

stole

carbonara

...

thing

matter

object

phenomenon

body part

body of water

headhair

veinmouth

ocean

cloud

snowwave

softmax

softmax

softmax softmax

softmax

softmax

WordTree1k

Imagenet 1k

986

1355

Figure 5: Prediction on ImageNet vs WordTree. Most Ima-geNet models use one large softmax to predict a probability distri-

bution. Using WordTree we perform multiple softmax operations

over co-hyponyms.

Dataset combination with WordTree. We can use

WordTree to combine multiple datasets together in a sen-

sible fashion. We simply map the categories in the datasets

to synsets in the tree. Figure 6 shows an example of using

WordTree to combine the labels from ImageNet and COCO.

WordNet is extremely diverse so we can use this technique

with most datasets.

Joint classification and detection. Now that we can

combine datasets using WordTree we can train our joint

model on classification and detection. We want to train

an extremely large scale detector so we create our com-

bined dataset using the COCO detection dataset and the

top 9000 classes from the full ImageNet release. We also

need to evaluate our method so we add in any classes from

the ImageNet detection challenge that were not already in-

cluded. The corresponding WordTree for this dataset has

9418 classes. ImageNet is a much larger dataset so we bal-

ance the dataset by oversampling COCO so that ImageNet

is only larger by a factor of 4:1.

Using this dataset we train YOLO9000. We use the base

YOLOv2 architecture but only 3 priors instead of 5 to limit

the output size. When our network sees a detection image

we backpropagate loss as normal. For classification loss, we

only backpropagate loss at or above the corresponding level

of the label. For example, if the label is “dog” we do assign

any error to predictions further down in the tree, “German

Shepherd” versus “Golden Retriever”, because we do not

have that information.

When it sees a classification image we only backpropa-

gate classification loss. To do this we simply find the bound-

ing box that predicts the highest probability for that class

7269

airplane apple backpack banana bat bear bed bench bicycle bird

.....zebra70

COCO

Afghanhound

Africanchameleon

Africancrocodile

Africanelephant

Africangrey

Africanhunting dog

Airedale Americanalligator

Americanblack bear

Americanchameleon

.....zucchini22k

ImageNet

animal artifact natural object phenomenon

plantfungusvehicle

equipmentcat

dog fish

tabby Persian

ground water air

airplanecar

biplane jet airbus stealthfighter

houseplant

vascularplant

physical object WordTree

goldenfern

potatofern

feltfern

sealavender

Americantwinflower

Figure 6: Combining datasets using WordTree hierarchy. Us-ing the WordNet concept graph we build a hierarchical tree of vi-

sual concepts. Then we can merge datasets together by mapping

the classes in the dataset to synsets in the tree. This is a simplified

view of WordTree for illustration purposes.

and we compute the loss on just its predicted tree. We also

assume that the predicted box overlaps what would be the

ground truth label by at least .3 IOU and we backpropagateobjectness loss based on this assumption.

Using this joint training, YOLO9000 learns to find ob-

jects in images using the detection data in COCO and it

learns to classify a wide variety of these objects using data

from ImageNet.

We evaluate YOLO9000 on the ImageNet detection task.

The detection task for ImageNet shares on 44 object cate-

gories with COCO which means that YOLO9000 has only

seen classification data for the majority of the test cate-

gories. YOLO9000 gets 19.7 mAP overall with 16.0 mAP

on the disjoint 156 object classes that it has never seen any

labelled detection data for. This mAP is higher than results

achieved by DPM but YOLO9000 is trained on different

datasets with only partial supervision [4]. It also is simulta-

neously detecting 9000 other categories, all in real-time.

YOLO9000 learns new species of animals well but strug-

gles with learning categories like clothing and equipment.

New animals are easier to learn because the objectness pre-

dictions generalize well from the animals in COCO. Con-

versely, COCO does not have bounding box label for any

type of clothing, only for person, so YOLO9000 struggles to

model categories like “sunglasses” or “swimming trunks”.

diaper 0.0

horizontal bar 0.0

rubber eraser 0.0

sunglasses 0.0

swimming trunks 0.0

...

red panda 50.7

fox 52.1

koala bear 54.3

tiger 61.0

armadillo 61.7

Table 7: YOLO9000 Best and Worst Classes on ImageNet.The classes with the highest and lowest AP from the 156 weakly

supervised classes. YOLO9000 learns good models for a variety of

animals but struggles with new classes like clothing or equipment.

5. Conclusion

We introduce YOLOv2 and YOLO9000, real-time de-

tection systems. YOLOv2 is state-of-the-art and faster

than other detection systems across a variety of detection

datasets. Furthermore, it can be run at a variety of image

sizes to provide a smooth tradeoff between speed and accu-

racy.

YOLO9000 is a real-time framework for detection more

than 9000 object categories by jointly optimizing detection

and classification. We use WordTree to combine data from

various sources and our joint optimization technique to train

simultaneously on ImageNet and COCO. YOLO9000 is a

strong step towards closing the dataset size gap between de-

tection and classification.

Many of our techniques generalize outside of object de-

tection. Our WordTree representation of ImageNet offers a

richer, more detailed output space for image classification.

Dataset combination using hierarchical classification would

be useful in the classification and segmentation domains.

Training techniques like multi-scale training could provide

benefit across a variety of visual tasks.

For future work we hope to use similar techniques for

weakly supervised image segmentation. We also plan to

improve our detection results using more powerful match-

ing strategies for assigning weak labels to classification data

during training. Computer vision is blessed with an enor-

mous amount of labelled data. We will continue looking

for ways to bring different sources and structures of data

together to make stronger models of the visual world.

Acknowledgements: We would like to thank Junyuan Xie

for helpful discussions about constructing WordTree. This

work is in part supported by ONR N00014-13-1-0720, NSF

IIS-1338054, NSF-1652052, NRI-1637479, Allen Distin-

guished Investigator Award, and the Allen Institute for Ar-

tificial Intelligence.

7270

References

[1] S. Bell, C. L. Zitnick, K. Bala, and R. Girshick. Inside-

outside net: Detecting objects in context with skip

pooling and recurrent neural networks. arXiv preprint

arXiv:1512.04143, 2015. 6

[2] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-

Fei. Imagenet: A large-scale hierarchical image database.

In Computer Vision and Pattern Recognition, 2009. CVPR

2009. IEEE Conference on, pages 248–255. IEEE, 2009. 1

[3] M. Everingham, L. Van Gool, C. K. Williams, J. Winn, and

A. Zisserman. The pascal visual object classes (voc) chal-

lenge. International journal of computer vision, 88(2):303–

338, 2010. 1

[4] P. F. Felzenszwalb, R. B. Girshick, and D. McAllester.

Discriminatively trained deformable part models, release 4.

http://people.cs.uchicago.edu/ pff/latent-release4/. 8

[5] R. B. Girshick. Fast R-CNN. CoRR, abs/1504.08083, 2015.

5, 6

[6] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learn-

ing for image recognition. arXiv preprint arXiv:1512.03385,

2015. 2, 5, 6

[7] S. Ioffe and C. Szegedy. Batch normalization: Accelerating

deep network training by reducing internal covariate shift.

arXiv preprint arXiv:1502.03167, 2015. 2, 5

[8] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet

classification with deep convolutional neural networks. In

Advances in neural information processing systems, pages

1097–1105, 2012. 2

[9] M. Lin, Q. Chen, and S. Yan. Network in network. arXiv

preprint arXiv:1312.4400, 2013. 4

[10] T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ra-

manan, P. Dollár, and C. L. Zitnick. Microsoft coco: Com-

mon objects in context. In European Conference on Com-

puter Vision, pages 740–755. Springer, 2014. 1, 6

[11] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, and S. E. Reed.

SSD: single shot multibox detector. CoRR, abs/1512.02325,

2015. 5, 6

[12] G. A. Miller, R. Beckwith, C. Fellbaum, D. Gross, and K. J.

Miller. Introduction to wordnet: An on-line lexical database.

International journal of lexicography, 3(4):235–244, 1990.

6

[13] J. Redmon. Darknet: Open source neural networks in c.

http://pjreddie.com/darknet/, 2013–2016. 5

[14] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi. You

only look once: Unified, real-time object detection. arXiv

preprint arXiv:1506.02640, 2015. 5, 6

[15] S. Ren, K. He, R. Girshick, and J. Sun. Faster r-cnn: To-

wards real-time object detection with region proposal net-

works. arXiv preprint arXiv:1506.01497, 2015. 2, 3, 5, 6

[16] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh,

S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein,

A. C. Berg, and L. Fei-Fei. ImageNet Large Scale Visual

Recognition Challenge. International Journal of Computer

Vision (IJCV), 2015. 2

[17] K. Simonyan and A. Zisserman. Very deep convolutional

networks for large-scale image recognition. arXiv preprint

arXiv:1409.1556, 2014. 2, 4

[18] C. Szegedy, S. Ioffe, and V. Vanhoucke. Inception-v4,

inception-resnet and the impact of residual connections on

learning. CoRR, abs/1602.07261, 2016. 2

[19] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed,

D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich.

Going deeper with convolutions. CoRR, abs/1409.4842,

2014. 4

[20] B. Thomee, D. A. Shamma, G. Friedland, B. Elizalde, K. Ni,

D. Poland, D. Borth, and L.-J. Li. Yfcc100m: The new

data in multimedia research. Communications of the ACM,

59(2):64–73, 2016. 1

7271

http://pjreddie.com/darknet/